Semi-adaptive non-intrusive electric load monitoring method based on deep learning

Abstract

The application proposes a semi-adaptive non-intrusive electrical load monitoring method based on deep learning, which constructs a data set through sampling, overlapping window segmentation, appliance activation signal extraction, and random combination of activation signals of the load signal obtained from the household electric meter. By designing a neural network model containing a separation main network and an individual feature extraction sub-network, the steps of extracting general and individual features of appliances are modularized and specified. Through deep learning in a multi-task mode, the functions of different modules are constrained, and an optimized deep learning model is obtained. The application effectively solves the problem of insufficient generalization and universality of the non-intrusive electrical load decomposition method based on deep learning, and the method has semi-adaptive monitoring capability for different types of individual appliances.

Description

Technical Field

[0001] This invention relates to the field of power load monitoring technology, and in particular to a semi-adaptive, non-intrusive power load monitoring method based on deep learning. Background Technology

[0002] With economic development and improved living standards, residential electricity consumption is steadily increasing annually, and this rapid increase in energy consumption exacerbates negative environmental impacts. Monitoring the energy consumption of household appliances can help policymakers and users understand the composition, patterns, and characteristics of household energy consumption, thereby assisting policymakers in formulating energy-saving and emission-reduction policies and guiding users to rationally allocate their electrical equipment usage. Traditional Intrusive Load Monitoring (ILM) requires each electrical device to be equipped with a measurement device with communication capabilities, which increases the cost of deploying and maintaining measurement equipment. Non-Intrusive Load Monitoring (NILM), on the other hand, only requires the installation of monitoring instruments at the point of entry into the power grid. Algorithms are used to analyze the collected total electricity consumption information to monitor the electricity consumption status of each electrical device. For large-scale deployments, NILM systems can significantly reduce installation complexity and maintenance costs.

[0003] To date, the most widely studied non-invasive electrical load monitoring methods fall into two main categories: the traditional method combining Markov models with sequential Bayesian inference, and the deep learning-based method. Currently, the deep learning-based method, supported by data-driven approaches, far surpasses the traditional method in both accuracy and efficiency.

[0004] However, current deep learning-based methods also have their problems: they rely too much on specific training datasets, resulting in insufficient generalization ability and poor universality; ordinary deep learning-based non-invasive electrical load monitoring methods often only have good monitoring effects on the types and models of appliances included in the dataset, while performing poorly on other types and models of appliances outside the dataset. This is because these neural networks learn more about the common features of various types, such as the common features of different models of refrigerators, while ignoring the specific features of different models of the same type of appliance. For example, even among refrigerators, different brands of refrigerators will have very different power curves when working, and these specific features play an extremely important role in load monitoring. Summary of the Invention

[0005] The main objective of this invention is to overcome the shortcomings of existing deep learning-based non-intrusive electrical load monitoring methods and propose a semi-adaptive non-intrusive electrical load monitoring method based on deep learning. This method organically combines the feature pattern extraction of individual target electrical appliances with the decomposition of electrical load signals, effectively solving the problem of insufficient generalization ability for different types of electrical appliances in existing deep learning-based load monitoring methods.

[0006] The present invention adopts the following technical solution:

[0007] A semi-adaptive, non-intrusive electrical load monitoring method based on deep learning is characterized by the following steps:

[0008] 1) Establish a total dataset using total user load and individual appliance load information;

[0009] 2) Construct a deep learning model that includes a separate main network and a feature extraction sub-network;

[0010] 3) Set the relevant training parameters and train the network to obtain the optimized deep learning model;

[0011] 4) Network testing: The optimized deep learning model is used to monitor the load of target electrical appliances according to user needs.

[0012] Preferably, in step 1), the total user load and the individual appliance load information are sampled, segmented by overlapping windows, appliance activation signals are extracted, and the activation signals are recombined to obtain a total dataset; the data in the total dataset are divided into three categories: total meter load signal, individual appliance load signals, and individual appliance activation signals; the mathematical model of a single sample in the total dataset can be expressed as:

[0013]

[0014] Where t represents time, y(t) represents the total household load signal at time t, and x i (t) represents the load signal of the i-th electrical appliance at time t, n represents the total number of electrical appliances in the system, and N(t) represents the noise signal.

[0015] Preferably, the sampling and overlapping window segmentation refers to: uniformly sampling the total user load and the single appliance load information over time to reduce data redundancy, and using overlapping windows of different sizes to segment the original long-time sequence signal into short-time sequences according to the different operating characteristics of each appliance.

[0016] Preferably, the extraction and recombination of the electrical appliance activation signal refers to: extracting x respectively iThe load signal of the electrical appliance in (t) when it is in working state is used as the activation signal of the electrical appliance. Then, the activation signals of different electrical appliances at different time periods are randomly combined and mixed to form new sample data.

[0017] Preferably, in step 2), the separation main network is constructed by stacking convolutional layers, including a front-end convolutional module and a back-end convolutional module. The front-end convolutional module extracts universal features of electrical appliances, and the back-end convolutional module performs load decomposition tasks for the target electrical appliances. The back-end convolutional module has two inputs: one input is the latent variable passed from the front-end convolutional module, and the other input is the feature information passed from the feature extraction sub-network. The back-end convolutional module performs feature merging processing on the two inputs.

[0018] Preferably, the feature extraction subnetwork includes convolutional layers, residual network modules, and fully connected layers; the feature extraction subnetwork takes the activation load signal of the target electrical appliance as input and outputs two things: one is the target signal features extracted from the average pooling layer output network, and the other is the result of the target electrical appliance type judgment output from the linear layer.

[0019] Preferably, in step 3), the deep learning model adopts a multi-task objective deep learning approach, simultaneously constraining the separation performance of the main separation network and the feature extraction performance of the feature extraction sub-network. The loss function is as follows:

[0020]

[0021]

[0022] L = L1 + L2

[0023] Where L1 is the mean square error between the output of the main network and the label, T is the total length of the load signal, t represents time, y(t) is the total load signal changing with time t, x(t) is the actual load signal of the target appliance changing with time t, w represents the network parameters, A represents the target appliance feature information extracted by the feature extraction subnetwork, f(·) represents the mapping relationship represented by the network model; L2 is the cross entropy between the output of the feature extraction subnetwork and the label, M represents the total number of appliance types, a m y represents the activation signal of the m-th type of electrical appliance. m Let g be the label of the m-th type of appliance, and g(·) represent the mapping relationship represented by the network model; L1 and L2 are added together to obtain the total loss function L for network training.

[0024] Preferably, in step 3), the input data during network training is divided into two categories: one category includes the total load signal and the target appliance activation signal, which is used to train the separation main network and the feature extraction sub-network; the other category includes various different appliance activation signals, which is used to train the feature extraction sub-network.

[0025] Preferably, the specific method for training the network is as follows:

[0026] Input the total dataset obtained in step 1) into the deep learning model to train the network. Use the backpropagation algorithm combined with the multi-task model for training, calculate the loss function, and update the parameters in the network accordingly. When the predetermined maximum number of iterations is reached or the loss function no longer decreases, terminate the training to obtain the optimized deep learning model.

[0027] Preferably, in step 4), the network test specifically involves: inputting the total dataset obtained in step 1) into the optimized deep learning model to obtain the target electrical load signal extracted by the network, and comparing it with the label to verify the performance of the optimized deep learning model.

[0028] As can be seen from the above description of the present invention, compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. The method of the present invention realizes the function of extracting the target appliance load signal from the total household electricity load signal mixed with multiple appliance load signals. By extracting and remixing the appliance activation signals from the raw data collected by the electricity meter, the amount of data is effectively expanded and the data diversity is increased, reducing the requirements for the size of the training dataset and enhancing the network generalization ability to a certain extent.

[0030] 2. The method of this invention, by establishing a neural network model combining a main network and a feature extraction sub-network, not only focuses on the network's ability to identify and separate the general characteristics of target type electrical appliances, but also pays attention to the network's ability to extract the unique features of different types of electrical appliances. Therefore, in non-intrusive electrical load monitoring, compared with traditional methods and other deep learning-based methods, the method proposed in this invention has a greater advantage in load monitoring accuracy. Furthermore, due to its adaptive feature extraction capability for different electrical appliances, it also significantly surpasses other methods in terms of generalization and universality. Attached Figure Description

[0031] Figure 1(a) shows the mixed total power signal, including the dishwasher power signal;

[0032] Figure 1(b) shows the dishwasher power signal, i.e., the tag, contained in Figure 1(a);

[0033] Figure 1(c) shows an activation signal for a dishwasher;

[0034] Figure 2(a) shows the overall network architecture;

[0035] Figure 2(b) shows the structure of the stacked convolutional CNNs used in Figure 2(a);

[0036] Figure 2(c) shows the structure of the residual network module (ResBlock) used in Figure 2(a);

[0037] Figure 3(a) shows the separation results using a dishwasher as an example. From left to right, the results are the total power signal, the tag, and the network separation.

[0038] Figure 3(b) shows the separation results using a kettle as an example. From left to right, the results are the total power signal, the tag, and the network separation.

[0039] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Detailed Implementation

[0040] The present invention will be further described below through specific embodiments.

[0041] In this invention, the terms "first," "second," and "third," etc., are used only to distinguish similar objects and are not necessarily used to describe a specific order or sequence, nor should they be construed as indicating or implying relative importance. The use of terms such as "upper," "lower," "left," "right," "front," and "rear" to indicate orientation or positional relationships is based on the orientation or positional relationships shown in the accompanying drawings and is only for the convenience of describing the invention, not to indicate or imply that the device referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation on the scope of protection of this invention. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0042] Furthermore, in the description of this application, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0043] A semi-adaptive, non-intrusive electrical load monitoring method based on deep learning includes the following steps:

[0044] 1) Establish a total dataset by combining the total load of users with the load information of individual appliances.

[0045] The total user load and individual appliance load information are sampled, segmented using overlapping windows, and appliance activation signals are extracted and recombined to obtain the total dataset. The data in the total dataset are then divided into three categories: total meter load signal, individual appliance load signals, and individual appliance activation signals. The mathematical model for a single sample in the total dataset can be expressed as:

[0046]

[0047] Where t represents time, y(t) represents the total household load signal at time t, and x i (t) represents the load signal of the i-th electrical appliance at time t, n represents the total number of electrical appliances in the system, and N(t) represents the noise signal.

[0048] Sampling and overlapping window segmentation refers to uniformly sampling the total user load and individual appliance load information over time to reduce data redundancy, and using overlapping windows of different sizes to segment the original long-term time series signal into short-term time series based on the different operating characteristics of each appliance.

[0049] Appliance activation signal extraction and activation signal recombination refer to: extracting x separately i The load signal of the electrical appliance in (t) when it is in working state is used as the activation signal of the electrical appliance. Then, the activation signals of different electrical appliances at different time periods are randomly combined and mixed to form new sample data.

[0050] Taking the open-source database UKDALE as an example, to ensure the network's generalization ability, different household-based datasets from UKDALE were selected for both the training and testing sets. The original power signals were then sampled at 6-second intervals. The sampled signals were then segmented using overlapping windows of size T and stride S, yielding the network's total input power signal and the power signals of multiple corresponding appliances. Simultaneously, the operating state power signals of each appliance in each household were extracted and used as activation signals for the input feature sub-network. Figure 1(a)-Figure 1(c) Examples of signals concentrated in the total data set are given. Figure 1(a) shows the total power signal; Figure 1(b) shows the target appliance power signal contained within the total power signal, i.e., the label; Figure 1(c) shows the target appliance activation power signal. Note that in the same sample, the label and the activation signal are not the same signal segment.

[0051] To expand the dataset, the activation signals of each appliance were randomly recombine and summed to obtain new sample data. The window size T and stride S of the overlapping window were varied depending on the characteristics of the target appliances. The total power signal of the households in the dataset was used as the input to the main separation network, the activation signals of each appliance in the household were used as the input to the feature extraction network, and the actual power signals of each appliance in the household were used as labels for network training, validation, and testing, thus constructing training, validation, and test sets.

[0052] 2) Build a deep learning model that includes a separate main network and a feature extraction sub-network.

[0053] The main network is constructed using stacked convolutional layers, including front-end and back-end convolutional modules. The front-end convolutional module extracts general features of electrical appliances, while the back-end convolutional module performs load decomposition of the target electrical appliances. The back-end convolutional module has two inputs: one is the latent variable passed from the front-end convolutional module, and the other is the feature information passed from the feature extraction subnetwork. The back-end convolutional module performs feature merging processing on the two inputs.

[0054] The feature extraction subnetwork consists of convolutional layers, residual network modules, and fully connected layers. The feature extraction subnetwork takes the activation load signal of the target appliance as input and outputs two things: one is the target signal features extracted from the average pooling layer output network, and the other is the result of the target appliance type judgment output from the linear layer.

[0055] Figure 2(a) shows the overall network model, which includes the main structure of the network model, the loss function, and the relevant training parameters. The main structure of the network model includes a segregating main network and a feature extraction sub-network. The segregating main network consists of one convolutional layer, one 1×1 convolutional layer, B convolutional stacking modules, one 1×1 convolutional layer, one attention layer, and one convolutional layer. The convolutional stacking module, as shown in Figure 2(b), consists of one 1×1 convolutional layer, one dilated convolutional layer, and one 1×1 convolutional layer. Note that this module has two inputs: one is the latent variable passed from the previous neural network layers, and the other is the feature information passed from the feature extraction module. The module performs feature merging processing on the two inputs. The feature extraction sub-network, in the order of input to output, consists of one convolutional layer, one 1×1 convolutional layer, R residual network modules, one 1×1 convolutional layer, one average pooling layer, and one linear layer. The residual network module, as shown in Figure 2(c), consists of two 1×1 convolutional layers and one max pooling layer. The input of this module is the activation power signal of the target electrical appliance, and the output has two parts: one is the target signal feature extracted from the average pooling layer output network, and the other is the result of the target electrical appliance type judgment output from the linear layer.

[0056] 3) Set the relevant training parameters and train the network to obtain the optimized deep learning model.

[0057] The deep learning model employs a multi-task objective approach, simultaneously constraining the separation performance of the main network and the feature extraction performance of the feature extraction sub-network. The loss function used during network training is shown below:

[0058]

[0059]

[0060] L = L1 + L2

[0061] Where L1 is the mean square error between the output of the main network and the label, T is the total length of the load signal, is the total load signal, is the actual load signal of the target appliance, t represents time, y(t) is the total load signal changing with time t, x(t) is the actual load signal of the target appliance changing with time t, w represents the network parameters, A represents the target appliance feature information extracted by the feature extraction subnetwork, f(·) represents the mapping relationship represented by the network model; L2 is the cross-entropy between the output of the feature extraction subnetwork and the label, M represents the total number of appliance types, a m y represents the activation signal of the m-th type of electrical appliance. m Let g be the label of the m-th type of appliance, and g(·) represent the mapping relationship represented by the network model; L1 and L2 are added together to obtain the total loss function L for network training.

[0062] During network training, the input data is divided into two categories: one category contains the total load signal and the target appliance activation signal, used to train the main separation network and the feature extraction subnetwork; the other category contains activation signals from various appliances, used to train the feature extraction subnetwork. The specific method for network training is as follows:

[0063] The total dataset obtained in step 1) is input into the deep learning model to train the network. The network parameters are adjusted until the multi-task training loss function decreases to convergence and tends to stabilize, resulting in a functional network model. The total load signal from the training set is input into the main network, and the target appliance activation signal is input into the feature extraction sub-network. The network is trained using a backpropagation algorithm combined with multi-task model training, calculating the loss function, and updating the network parameters accordingly. Training is terminated when the predetermined maximum number of iterations is reached or the loss function no longer decreases, resulting in a functional network model.

[0064] 4) Network testing: The optimized deep learning model is used to monitor the load of target electrical appliances according to user needs.

[0065] The network test is as follows: the total power signal in the total dataset obtained in step 1) is input into the separation main network of the optimized deep learning model obtained in step 3), the target appliance activation signal is input into the feature extraction sub-network, the target appliance power signal separated by the network is obtained, and it is compared with the label generated in step 1) to verify the performance of the optimized deep learning model.

[0066] Figures 3(a) and 3(b) demonstrate examples of separating the target appliance power signal from the total power signal using a semi-adaptive, non-intrusive load decomposition method based on multi-task mode deep learning. Figure 3(a) shows the separation result using a dishwasher as an example, with the total power signal, tag, and network separation results from left to right; Figure 3(b) shows the separation result using a kettle as an example, with the total power signal, tag, and network separation results from left to right.

[0067] Since the sample households in the test set are different from those in the training set, the functionalized network has strong generalization ability and performs well on the test set, effectively extracting the power signal of the target appliance from the total electricity meter signal.

[0068] The above are merely specific embodiments of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention using this concept shall be considered as infringing upon the protection scope of the present invention.

Claims

1. A semi-adaptive, non-invasive electrical load monitoring method based on deep learning, characterized in that: Includes the following steps: 1) Establish a total dataset using user total load and individual appliance load information; sample the user total load and individual appliance load information, segment using overlapping windows, extract appliance activation signals, and recombine the activation signals to obtain the total dataset; divide the data in the total dataset into three categories: total meter load signal, individual appliance load signals, and individual appliance activation signals; the mathematical model of a single sample in the total dataset is expressed as: ； in Indicates time, express Moment-time sample of total household load signal. express Time of the first The load signal of an appliance, where n represents the total number of appliances in the system. Indicates a noise signal; The extraction and recombination of the electrical appliance activation signal refer to: extracting the activation signal and recombining the activation signal respectively. The load signal of the electrical appliance when it is in working state is used as the activation signal of the appliance. Then, the activation signals of different appliances at different time periods are randomly combined and mixed to form new sample data. 2) Construct a deep learning model that includes a separate main network and a feature extraction sub-network; The separation main network is constructed by stacking convolutional layers, including a front-end convolutional module and a back-end convolutional module. The front-end convolutional module extracts universal features of electrical appliances, and the back-end convolutional module performs load decomposition of the target electrical appliances. The back-end convolutional module has two inputs: one is the latent variable passed from the front-end convolutional module, and the other is the feature information passed from the feature extraction sub-network. The back-end convolutional module performs feature merging processing on the two inputs. The feature extraction subnetwork includes convolutional layers, residual network modules, and fully connected layers. The feature extraction subnetwork takes the activation load signal of the target appliance as input and outputs two features: one is the target signal feature extracted from the average pooling layer output network, and the other is the result of the target appliance type judgment output from the linear layer. 3) Set the relevant training parameters and train the network to obtain the optimized deep learning model; The deep learning model employs a multi-task objective deep learning approach, simultaneously constraining the separation performance of the main separation network and the feature extraction performance of the feature extraction sub-network. The loss function used during network training is shown below: ； ； in, To separate the mean square error between the main network output and the label, The total length of the load signal. Indicates time, For time The changing total load signal For time The actual load signal of the changing target electrical appliance Represents network parameters, This represents the target electrical appliance feature information extracted by the feature extraction subnetwork. This represents the mapping relationship depicted by the network model; The cross-entropy is the ratio of the output of the feature extraction subnetwork to the label, where M represents the total number of appliance types. This represents the activation signal of the m-th type of electrical appliance. Let m be the label for the type of electrical appliance. This represents the mapping relationship depicted by the network model; and The sum is the total loss function for network training. ; 4) Network testing: The optimized deep learning model is used to monitor the load of target electrical appliances according to user needs.

2. The semi-adaptive, non-invasive electrical load monitoring method based on deep learning as described in claim 1, characterized in that: The sampling and overlapping window segmentation refer to: uniformly sampling the total user load and the single appliance load information over time to reduce data redundancy, and using overlapping windows of different sizes to segment the original long-term time series signal into short-term time series according to the different operating characteristics of each appliance.

3. The semi-adaptive, non-invasive electrical load monitoring method based on deep learning as described in claim 1, characterized in that: In step 3), the input data during network training is divided into two categories: one category includes the total load signal and the target appliance activation signal, which is used to train the separation main network and the feature extraction sub-network; the other category includes various different appliance activation signals, which is used to train the feature extraction sub-network.

4. The semi-adaptive, non-invasive electrical load monitoring method based on deep learning as described in claim 1, characterized in that: The specific method for training the network is as follows: Input the total dataset obtained in step 1) into the deep learning model to train the network. Use the backpropagation algorithm combined with the multi-task model for training, calculate the loss function, and update the parameters in the network accordingly. When the predetermined maximum number of iterations is reached or the loss function no longer decreases, terminate the training to obtain the optimized deep learning model.

5. The semi-adaptive, non-invasive electrical load monitoring method based on deep learning as described in claim 1, characterized in that: In step 4), the network test specifically involves inputting the total dataset obtained in step 1) into the optimized deep learning model to obtain the target electrical load signal extracted by the network, and comparing it with the label to verify the performance of the optimized deep learning model.

Images